On Architecture for the Future Petascale Computing
Ku ccaron era Lud ecaron k · Advances in parallel computing · 2018
The supercomputers listed in Top500 and running the HPCG benchmark are usually not able to use more than 1–5 % of their peak computing power, and no system is able to achieve 1 PFlop/s. The paper shows that the Conjugate Gradient algorithm without preconditioning requires only weak bandwidth of interconnect links, but, being a problem of low arithmetic intensity (flop/byte ratio), it would need the memory bandwidth about 20 times greater than available.